Drop-in persistent memory for your AI coding agent.
AI coding agents lose context between sessions. They re-learn your project every time, re-break the same things, forget why a decision was made. The fix isn't a smarter model. It's three habits: a boot read at the start of a session, a write-back at the end, and a git trail so the memory is durable and reviewable.
Mini Z is that, as a tiny folder of Markdown you drop into any project, wired into whatever agent you use.
npx github:svongripp/mini-z initNot on npm yet, so the
github:form above is the one that works today. Once it's published this shortens tonpx mini-z init.
It asks a few quick questions (which agent, how it should lean, project posture, tone), then:
- creates a
mini-z-memory/folder (your project's memory), and - adds a short, clearly-marked block to your agent's instructions file
(
CLAUDE.md,AGENTS.md,.github/copilot-instructions.md, …) telling it to read that memory at the start of every session and update it at the end.
Zero dependencies, nothing to install, ~30 seconds.
mini-z-memory/
README.md how the system works
project-north-star.md what this project is for (the compass)
gremlins.md gotchas that have bitten this project before
decisions.md append-only "why it's this way" log
worklog.md terse session log
And a marked block in your agent's rules file:
<!-- mini-z:start -->
## Mini Z — persistent project memory
Boot: read mini-z-memory/ … · Operating character: … · Write-back: …
<!-- mini-z:end -->
Re-running mini-z init updates that block in place and never touches the
rest of your instructions file.
| Question | Choices | Shapes |
|---|---|---|
| Agent | Claude Code · Codex · Cursor · Copilot · other | which instructions file gets the block |
| Mode | adaptive · architect · executor | how the agent leans (plan vs. build) |
| Posture | personal/experimental · production | how cautious it is (read-only-by-default for prod) |
| Tone | neutral · terse · warm | how it communicates |
| Project | one line (optional) | seeds the north-star |
The core is plain Markdown, so it works with Claude Code, Codex, Cursor, Copilot, or anything that reads a project instructions file. Only the filename differs; Mini Z picks the right one. Claude Code users can optionally add hooks for automatic boot.
It started as the executor half of a personal AI setup — the worker that keeps its head in the codebase while a separate partner holds the strategy. This is that memory layer, generalized so anyone can drop it into their own project.
MIT.